I’ve been meaning to blog about this paper, because its abstract certainly promises a lot:
. . .Target prediction software based on machine learning models correctly identified additional macromolecular targets of the computationally designed compound and the structurally related marketed drug azosemide. The study validates computational de novo design as a prime method for generating chemical probes and starting points for drug discovery.
They’re looking at inhibition of the kinase DAPK3, which has a wide variety of physiologic roles, and is a potential target in several disease states. It’s known to have some similarities to ROCK1, for which there’s a marketed inhibitor (fasudil) that’s pretty much fragment-sized itself. The compound is also a micromolar hit on DAPK3, so it’s a perfect starting point. The team used their already-reported “Design of Genuine Structures” (DOGS) software to generate plausible fragment-sized analogs of this compound’s structure and presumed pharmacophore, then used a machine-learning program to try to prioritize these fragments for estimated activity on DAPK3.
One of the top-ranking hits was 3-(tetrazoyl) phenylsulfonamide, and this one did indeed have high ligand efficiency on DAPK3. (No word on what the other top-ranking compounds did against the target, though, and I would actually be curious about that). They found that this fragment was surprisingly selective against a kinase panel, and used further runs of their prediction software to see what its other targets might be. Carbonic anhydrase came up, but you don’t need fancy software for that: any primary arylsulfonamide has to be presumed as a carbonic anhydrase inhibitor until proven otherwise.
An X-ray crystal structure with DAPK was obtained, and it showed that the tetrazole was a hinge-binding motif in the ATP binding site. As it turns out, there’s another marketed drug with a 3-tetrazolyl sulfonamide, a diuretic called azosemide. It also shows activity on DAPK3, which is apparently the first target that’s been assigned to the drug (although you’d also assume that it’s a carbonic anhydrase inhibitor as well). The authors finish up here by saying “The results of this study pinpoint the ability of ligand-based de novo design to generate synthesizable, fragment-like starting points that provide innovative chemical structures for “growth” into efficacious leads”.
Well, do they? The biggest leap I can see is from the ROCK1 inhibitor structure to the tetrazole sulfonamide core, and that’s a nice find. But there are a number of controls that I’d like to know more about. As mentioned earlier, there doesn’t seem to anything about the other compounds that were predicted to be as good (or better) than that one – a list of the predicted scaffolds is not provided, either in the paper or in its supporting information. It would be good to know what the hit rate was for the machine-learning program’s output. It would also be worth knowing what the hit rate of DAPK3 is against a plain-vanilla library of fragment structures – that is, what could you get by direct screening, versus the computational approach? Getting fragment hits for the ATP-binding site of a kinase should not be that hard, to be honest (in fact, here’s a crystal structure of a fragment in a DAPK binding site, from the first paper to ever report a small-molecule inhibitor against these kinases). I also would have expected a paper on computational approaches to a DAPK inhibitor to reference a five-year-old paper on virtual screening approaches to DAPK inhibitors, particularly when a compound from that latter paper is available commercially as a reference DAPK3 inhibitor. But apparently not. And as mentioned, I’m also not amazed that this paper’s compound and azosemide were both predicted to be carbonic anhydrase inhibitors; that’s not a computational triumph. So while I’m not discounting this approach, this paper doesn’t sell me on it, either. I’d like to see more before making up my mind.